As models improve, AI startups must prove they’re still necessary

Auf Deutsch lesen

10 min read

Also available in German: Bessere Modelle, weniger Vorsprung: Was KI-Start-ups noch unersetzlich macht.

Show Your Work

In one of our weekly sprint meetings, a simple question stalled everything. Part of the answer sat in the sales figures and the rest in customer records, held in two systems that didn’t talk to each other, so someone exported both and reconciled them in a spreadsheet while the meeting waited, and everyone treated that as normal.

It wasn’t a one-off. I have watched the same thing at two other companies in different industries, plenty of software each time, and still an ordinary question meant finding the person who knew where the data was buried. That’s when it stopped feeling like a bad week and started looking like a pattern.

Which brings me to Zeit AI. The Munich startup reached €1 million in annual recurring revenue within twelve months with a team of six, according to its Y Combinator profile. It connects scattered company systems and turns plain-language questions into reports and applications.

If you run or back a similar company, you already face the question. Which parts of what you sell stay yours once the underlying model can do it too? The revenue says companies will pay to avoid exports, spreadsheets and the hunt for someone who understands every field. It doesn’t say what survives better models writing SQL, reading databases and generating reports.

Zeit’s potential defence sits around what it learns and builds each time it connects to a customer’s systems. That only becomes a moat if it makes the next customer cheaper to serve. If it doesn’t, the recurring revenue is hiding a business that runs on custom engineering, not software.

Across the AI market, technical progress can erase scarcity, force a move up the stack or increase demand for the infrastructure around it. Coqui, Jasper, Aleph Alpha, n8n and Chift show those routes.

Exhibits A–F

Scarcity Left Early

Berlin-based Coqui built open-source text-to-speech and voice-cloning systems alongside a commercial web application and API. Founded in 2021 by engineers from Mozilla’s speech team, it built models such as XTTS, which could generate a voice in many languages from a short sample, and Coqui Studio was the attempt to turn that adoption into paying customers.

On January 3, 2024, co-founder Josh Meyer announced the shutdown. He didn’t name one cause. Stanford researcher Christopher Manning described standalone text-to-speech as a difficult business to sustain, and contemporary reporting pointed to funding and monetisation problems, while larger laboratories and open-source projects were making convincing speech generation more widely available. The capability was getting less scarce as the business struggled, and the technology stayed available after the company closed.

Jasper started from a stronger position. Its early product generated advertising copy, blog posts and other content from a prompt, and by October 2022 it had paying customers and a $1.5 billion valuation. Then ChatGPT launched the following month and put broadly similar writing capability inside a general-purpose chat window. Jasper offered more than basic text generation, but one of its central features was suddenly available everywhere.

After cutting roles in 2023, Jasper focused on marketing teams, brand workflows and enterprise integrations, which moved its value away from generic writing and toward workflow and customer context. I read that as the quiet admission.

Aleph Alpha faced pressure one layer lower. Founded in Heidelberg in 2019, it developed foundation models and was presented as Europe’s answer to OpenAI, and in 2023 it announced a financing package exceeding $500 million, including equity, research funding and commercial commitments.

OpenAI, Google and other larger laboratories kept improving their models while spending at a scale few European startups could match. As that standard rose, Aleph Alpha shifted towards PhariaAI, a platform for regulated companies and governments. Sifted reported that the change followed missed revenue targets and questions about the economics of remaining an independent model developer. Founder Jonas Andrulis stepped down as chief executive in October 2025, and the company agreed to a Cohere deal in April 2026, reportedly valuing the combined company at about $20 billion.

Aleph Alpha sits at a different layer from Zeit, but it shows the same risk at greater expense. A technical lead can shrink faster than the business adapts.

Boring Wins Quietly

Customers pay Zeit to write SQL, clean exports and build dashboards. Models are getting better at exactly that work, and established software companies can add similar features to products customers already use, so as those tasks get easier to reproduce, the harder-to-copy work starts at the connection to a live business.

A deployment can reveal SAP field meanings, finance-team naming habits, dashboard permissions and exceptions buried in historical data. None of that is glamorous, and all of it has to be connected to systems and access rules before software can act on it. Better models could make that knowledge more valuable, because they can do more with it once it’s available.

Zeit uses Forward-Deployed Engineers for this work. They may create reusable connectors and patterns that lower the cost of later deployments, or they may leave every account dependent on bespoke engineering. A repeatable product gets cheaper to deliver as it grows, while a service-heavy one can still be a good business, but each bit of revenue brings more implementation work with it.

Engines Need Roads

Better AI doesn’t weaken every business around it. More capable agents can create more activity that needs connecting, executing and governing.

Berlin-based n8n connects business applications and runs workflows across them. According to company-reported figures, revenue grew tenfold and users sixfold in 2025, while AI agents appeared in more than 80% of its workflows, and the company raised $180 million in October 2025 at a reported $2.5 billion valuation.

Founder Jan Oberhauser told the Raise Summit in Paris that the model is the engine, and everything n8n builds is “the vehicle, the streets, the rules” that let the engine take you somewhere. I like it because it’s the counterpoint in one sentence: a stronger engine can increase demand for the system around it.

Brussels-based Chift built one API connecting software companies to more than 120 European accounting, invoicing, payment and point-of-sale systems. It reported more than tenfold revenue growth between its 2024 seed round and a €10.5 million Series A in September 2026. Founder Gauthier Henroz argued that AI and e-invoicing were making interoperability more important as businesses adopted more tools, and the logic holds up: an AI system may decide an invoice needs reconciling, but it still needs authenticated, maintained access to the accounting software to do anything about it.

That defence isn’t permanent. Integrations can be rebuilt, customer context may become portable and established platforms can add orchestration.

Cheap Gets Crowded

Suppose leading models become much more capable and cheaper. Coding, reasoning and tool use all improve. A startup then has to ask which parts of its product customers will still need to buy from it, and the honest answer can be uncomfortable.

For Zeit, better models could cut the cost of generating a dashboard while raising the value of reliable access to customer systems and business context. For n8n, stronger agents may create more demand for workflows that carry out their decisions. Public figures don’t yet separate those effects.

Four Receipts, Please

• Deployment time: whether the next customer reaches a working application faster than the last.
• Engineering intensity: whether implementation hours grow more slowly than recurring revenue.
• Retention and expansion: whether customers keep the product and add more workflows after the first use case.
• Reusable revenue: how much growth comes from repeatable software rather than new custom projects.

Public evidence doesn’t show whether implementation time or engineering work per customer is falling, and that’s the difference between a useful service and a compounding software business. Zeit’s €1 million establishes demand. These four measures would show whether work inside one account makes the next one easier to win and serve.

Better underlying technology can strengthen a product while making its visible feature easier to copy. Durable value has to compound somewhere else.

For the buying decision behind orchestration, our n8n versus Make comparison sets out pricing and hosting assumptions. The IC Lab comparisons cover other agent and workflow categories.

Homework, Attached

This analysis draws on company profiles and announcements, funding disclosures, executive statements and contemporary reporting linked throughout the article. Company-reported growth figures are identified as such. The public evidence does not establish that foundation-model progress alone caused Coqui’s closure, Jasper’s repositioning or Aleph Alpha’s strategic shift. No company named in this article was contacted for comment.

Fair Questions

Spotted an error? Contact the desk. Our corrections policy explains how we handle corrections and updates.